How do I know which pages on my site ChatGPT is actually reading and linking to?
How to find which pages ChatGPT actually reads and links to: crawler logs, OAI-SearchBot and ChatGPT-User, citation tracking, referral traffic and joining tools.
Two separate questions hide in this one. "Reading" means ChatGPT's systems fetched the page. "Linking to" means ChatGPT showed the page as a source in an answer. A page can be read and never linked, and a page can be linked weeks after it was read. You need evidence for each, and then a way to join them.
This guide covers the evidence you can gather for free, the gaps in it, and the tools that put it together.
Part one: which pages is ChatGPT reading?
The evidence is in your server or CDN logs. OpenAI runs more than one crawler, and they mean different things:
- OAI-SearchBot is the bot associated with ChatGPT search results.
- GPTBot is associated with model training.
- ChatGPT-User appears when a user's request makes ChatGPT fetch a page on demand.
Read OpenAI's current crawler documentation to confirm each bot's purpose before you act on a log line, because roles can change.
What to do with logs
- Filter requests whose user agent contains these names.
- Group by URL and status code.
- Look for pages that are fetched often, pages never fetched and pages that return errors.
- Compare against your sitemap to find important pages that no bot has touched.
Treat user agents with caution. Anyone can claim to be a given bot. Real verification means checking that the source IP belongs to the operator, which most log tools do not do for you.
The robots.txt check
If a bot never appears, read your robots.txt and your CDN or firewall rules. A rule meant for scrapers often blocks AI crawlers too. Test the page as the bot would request it, then look again in the logs a few days later.
Part two: which pages is ChatGPT linking to?
Three sources of evidence, each partial.
Manual prompts. Ask ChatGPT the questions your page should answer and look at the sources shown. It is free and gives the most direct evidence, but it is a single run, and answers vary by wording, location and day.
Referral traffic. In GA4 or any analytics tool, look for visits from chatgpt.com. The landing pages with visits are pages that were linked and clicked. That undercounts, because linked pages that nobody clicked leave no trace, and some visits arrive without a referrer.
Browser tools. Promptwatch publishes a free GEO Inspector Chrome extension that shows ChatGPT's query fan-outs and which sources were fetched versus cited, and it runs locally. It is useful for checking a handful of prompts yourself.
None of these scales to hundreds of prompts, and none connects reading to linking.
Part three: join them
This is where a platform pays for itself. Promptwatch covers both sides:
Reading. Agent Analytics provides real-time, IP-verified crawler logs, with spoofed user agents filtered out. It tracks GPTBot, OAI-SearchBot and ChatGPT-User along with ClaudeBot, PerplexityBot and others. You get error tracking with failure explanations, stackable path filters and CSV export.
Linking. Citation analytics show page-level and domain-level citations with URL positions. Page Tracker follows specific URLs, yours or a competitor's. Citation Trends classifies each citation by content type and source type, and reports first cited date, ramp-up days, peak rate and decay half-life per page.
Joining. The crawl-to-citation path gives each page a citation rate. The 360 Insights view puts crawler logs, citation clicks and AI-referred visits per page side by side. Visitor analytics add conversions.
That lets you sort pages into four useful groups: read and cited, read and not cited, cited but rarely read (probably a cached or older version in use), and neither.
Logs reach it through Cloudflare (Logpush or an auto-deployed Worker), AWS CloudFront, Fastly, Vercel, Netlify, Akamai, Google Cloud CDN, a custom HTTP endpoint or the WordPress plugin. Crawler logs start on Professional ($245/mo, 25M logs) and are included on every agency plan from $199/mo. Essential has none.
Other tools to consider
| Tool | Reading | Linking | Price noted |
|---|---|---|---|
| Promptwatch | IP-verified crawler logs | Page-level citations, Page Tracker, Citation Trends | Professional $245/mo; agency from $199/mo |
| Scrunch AI | AI agent and bot traffic reporting | Visibility and citation tracking | Core $250/mo |
| Known Agents | Bot analytics with spoof verification | Estimated citations inferred from fetches | Free; $29/mo; $299/mo |
| Cloudflare AI Crawl Control | Requests per crawler and path, Cloudflare only | None | Free on every plan |
| AthenaHQ | Not listed | Citation and visibility tracking | Starter $295/mo |
Known Agents says its citation counts are inferred from fetches, which is not the same as observing a citation.
What to do with the answer
- Important page never read: fix access first.
- Read often, never cited: the content may be vague, thin or outranked by another source. Rewrite with a direct answer and specifics.
- Cited, but no visits: check the snippet and intent.
- Neither: decide if the page matters. If yes, link to it internally and make it easy to fetch.
A five-page starter audit
You do not need to analyze the whole site on day one. Pick the five pages that matter most to revenue. For each, search the logs for the three OpenAI bot names over the last month, ask ChatGPT the question the page should answer, and check analytics for chatgpt.com referrals. Write the result in one line per page: read or not, cited or not, visited or not. Patterns appear fast. Often the weak spot is the same across all five, such as a blocked path, a thin answer or a missing internal link.
A limit to keep in mind
Being read does not guarantee a link, and no setting guarantees that ChatGPT will cite you. The evidence shows where to look. It does not control the outcome.